Hive 处理count distinct 产生数据倾斜处理

问题描述

问题数据倾斜范畴，但是又不可以在Map端join，剔除特殊Key等方法进行处理。

set hive.groupby.skewindata=true;

insert overwrite table ad_overall_day partition(part_time='99', part_date='2015-11-99') 
select account_id, nvl(client_id,-1), nvl(track_id,'total'), sum(if(type=3,1,0)) as imp_cnt, 
sum(if(type=4,1,0)) as click_cnt, count(distinct if(type=3,zid,NULL)) as imp_uv,
count(distinct if(type=4,zid,NULL)) as click_uv 
from derived_di_v3 
where year='2015' and month='11' 
group by account_id, client_id, track_id 
grouping sets((account_id, client_id, track_id),(account_id, client_id),(account_id))   

But error find.
FAILED: SemanticException [Error 10022]: DISTINCT on different columns not supported with skew in data

separate group-by and join the results.

set hive.groupby.skewindata=true;
set hive.exec.parallel=true;

insert overwrite table ad_overall_day partition(part_time='99', part_date='2015-11-99') 
SELECT COALESCE(t1.account_id,t2.account_id),COALESCE(t1.client_id,t2.client_id),
    COALESCE(t1.track_id,t2.track_id),t1.imp_cnt,t1.imp_uv,t2.click_cnt,t2.click_uv
FROM
(select account_id, nvl(client_id,-1) as client_id, nvl(track_id,'total') as track_id, 
sum(if(type=3,1,0)) as imp_cnt, count(distinct if(type=3,zid,NULL)) as imp_uv
FROM derived_di_v3 where year='2015' and month='11' 
group by account_id, client_id, track_id 
grouping sets((account_id, client_id, track_id),(account_id, client_id),(account_id)) ) t1
FULL OUTER JOIN
(select account_id, nvl(client_id,-1) as client_id, nvl(track_id,'total') as track_id, 
sum(if(type=4,1,0)) as click_cnt, count(distinct if(type=4,zid,NULL)) as click_uv
FROM derived_di_v3 where year='2015' and month='11' 
group by account_id, client_id, track_id 
grouping sets((account_id, client_id, track_id),(account_id, client_id),(account_id)) ) t2  
ON t1.account_id = t2.account_id and t1.client_id = t2.client_id and t1.track_id = t2.track_id;

Cann’t run two MapReduce Job

Unfortunately, hive doesn’t explain the hql to two mapreduce job.
The parameter hive.groupby.skewindata seems has no affect.

Change the hql :

insert overwrite table ad_overall_day partition(part_time='99', part_date='2015-11-99')
select account_id, nvl(client_id,-1) as client_id, nvl(track_id,'total') as track_id, 
    sum(imp1) as imp_cnt,count(imp2) as imp_uv,sum(click1) as click_cnt,count(click2) as click_uv
FROM (select account_id, client_id, track_id,
     if(type=3,1,0) as imp1,if(type=3,zid,NULL) as imp2,
     if(type=4,1,0) as click1,if(type=4,zid,NULL) as click2
    FROM dmp.derived_di_v3 where year='2015' and month='11' 
    group by account_id, client_id, track_id,type,zid) t
group by account_id, client_id, track_id 
grouping sets((account_id, client_id, track_id),(account_id, client_id),(account_id))

Help Links

HIVE-474
HIVE-537